Argues that complex multi-agent tasks require explicit responsibility, handoff, and memory boundaries beyond joint optimization of agent and communication structures.
The concept of topological coherence posits that in complex multi-agent tasks, responsibility, handoff, and memory boundaries must remain consistent with task dependencies, a requirement not met by existing methods that only jointly optimize agent and communication structures. This principle is implemented in ToCoMAS (Topology-Coherent Multi-Agent System), which grounds task graphs in tool interfaces to derive reusable responsibility domains, dependency-aware collaboration, and boundary-regulated memory visibility. By ensuring these structural elements evolve together under explicit constraints, ToCoMAS improves task success, verified progress, and memory isolation compared to self-evolving baselines.
Topological Coherence for Self-evolving Multi-agent Systems frames self-evolving multi-agent systems as a structural, not merely behavioral, problem. The central argument is that jointly optimizing agent capabilities and communication topology is insufficient for complex, long-horizon tasks because such systems can become semantically incoherent: it becomes unclear which agent owns a decision, when control should transfer, and which memories are valid across subtasks or evolution steps. The paper introduces topological coherence as an organizing principle for keeping these systems legible and stable as they modify themselves.
The key insight is that multi-agent evolution should be constrained by explicit boundaries: responsibility boundaries that assign authority over subgoals or state, handoff boundaries that define when and how control or information transfers between agents, and memory boundaries that scope what persists, what is shared, and what is discarded. Rather than treating the agent graph as a free-form object to be optimized end-to-end, the work treats it as a topological structure whose invariants—such as ownership, interface continuity, and memory separation—should be preserved under mutation. This shifts the design focus from raw performance optimization toward maintaining a coherent division of labor and a predictable state-transition surface.
This matters because self-evolving multi-agent systems are increasingly likely to be used in open-ended, long-running settings where reliability, auditability, and composability are critical. Without explicit coherence constraints, evolution can produce ambiguous control loops, duplicated or conflicting responsibilities, and memory contamination across tasks. By making responsibility, handoff, and memory structure first-class objects, the paper offers a more principled route toward scalable agent systems that can improve over time without losing traceability, safety, or behavioral predictability.